Debiased Recommendation Based on Comparative Learning and Causal Embedding

Abstract Biases in recommendation systems significantly reduce recommendation accuracy and user experience. To address the issues in traditional recommendation systems: (1) inaccurate recommendation results due to ineffective modeling of users’ long-term and short-term interests, (2) popularity bias caused by the conformity effect, a debias recommendation method based on contrastive learning and causal embedding (CLACE). CLACE first employs two independent encoders to model users’ long-term and short-term interests separately. Then, a contrastive learning framework is designed to supervise the similarity between the long-term and short-term interest representations and interest proxies. A dynamic attention mechanism is introduced to adaptively adjust the weights of long-term and short-term interests to accurately reflect users’ preference biases. Simultaneously, to fundamentally reduce the popularity bias induced by the conformity effect during the recommendation process, a causal embedding model is introduced to separate user interests from the conformity effect, eliminating the negative impact of the conformity effect on recommendation results and achieving more precise recommendations. The effectiveness of the proposed CLACE is validated on two public datasets, and experimental results demonstrate that CLACE significantly improves recommendation accuracy, recall, and normalized discounted cumulative gain.